This post is also available in:
עברית (Hebrew)
Researchers at Tsinghua University in China developed a new photonic chip that processes, transmits, and reconstructs images in mere nanoseconds by skipping the optical to electronic data conversion deployed by conventional chips.
“Machine vision” is a developing field where cameras, sensors, and algorithms work together to make sense of the world around them and perform specific tasks. Up until now, technology relied on moving this data over long distances to analyze and take the appropriate response, but today’s fast-moving world demands that data is processed on the device (edge computing) to aid faster decision-making.
Lu Fang, associate professor at the Department of Electronic Engineering at Tsinghua University explained that the growth of edge devices (like smartphones, smart cars, laptops and more) resulted in an explosive growth of image data that needs to be processed, transmitted and displayed – “We are working to advance machine vision by integrating sensing and computing in the optical domain.”
“Capturing, processing and analyzing images for edge-based tasks such as autonomous driving is currently limited to millisecond-level speeds due to the necessity of optical-to-electronic conversions,” explained Fang. Nowadays, for computers to make sense of the data, images captured by machine vision devices need to be transferred from their optical nature to electronic versions.
According to Interesting Engineering, to solve this issue, the research team built an optical parallel computational array (OPCA) chip with a sensing-computing array made using ring resonators, a design that allows the photonic chip to convert an optical image into a two-dimensional representation of its light intensity that can be guided onto the chip using a micro-lens array. The chip has a processing bandwidth of up to a hundred billion pixels and a response time of just six nanoseconds.
Since the data is processed as light signals, the researchers used them to develop an all-optical neural network and deploy it for classification tasks typically carried out on the edge.
The research team used the chip for image classification of hand-drawn images and image convolution to demonstrate how the chip works. They share that in the future, they aim to increase the overall size of the OPCA chip and improve the neural network’s processing capacity to bring it closer to commercial usage.
